3 ms·
Not to put a damper on how cool this is (or bring up AI for the nth time), but I think GPT4 has mostly solved this problem, and you don't even need to learn a n
by johnfn 3y ago
Not to put a damper on how cool this is (or bring up AI for the nth time), but I think GPT4 has mostly solved this problem, and you don't even need to learn a new UI! For instance, GPT4 integrated into my IDE has a hit rate of close to 100% for translating english into code for simple things that the article mentions, such as creating new functions, moving arguments around, etc. (I feel silly using such a powerful tool for such simple tasks, but I often ask GPT4 to do tasks that I'm too lazy or don't want to hit all the keystrokes for, like converting an anonymous function to a lambda or whatever.) I suppose that the only difference is that it's not quite as fast, yet :)
- xena 3y agoThe latency will make or break this. I run all the models on-device as much as possible.
- Terr_ 3y agoPlus there's the whole privacy/autonomy aspect. Notwithstanding the last few decades of trend, many people still want certain data/behavior to stay firmly within a device they (supposedly) own and control, as opposed to owning a dumb-terminal.
- johnfn 3y agoI don't think that's true. Latency is critical when the limiting factor is that you know what you want to do but in order to do it, you need to issue many requests all at once. With GPT4 I can issue requests at a high enough level that the limiting factor is actually me thinking of what the next request will be.
- yjftsjthsd-h 3y agoI had a similar thought about ~AI helping, but AFAIK that still involves typing; is there a way to drive that via voice?
- kraig911 3y agoI'm seriously trying to do this but I keep struggling with integrating AI. I find it's constantly messing up or just not understanding what I'm asking. Mind going in depth about how you integrated GPT4? I'm using go-pilot and I recently got a GPT4 sub but I find myself writing a very long prompt over and over until I get a general idea of what I could do. Just for example I need to check for environment in js if it's prod, it did some statement like (environment === 'prod' || environment === 'prod-blue etc etc. I stripped it all down to just environment.includes('prod') lol. Another is just say if you're dealing with an api that returns a string for 'true' instead of true. It wrote a huge function about this...
- rpmisms 3y agoYou're probably trying to do too much with it at a time. I use GPT-4 to write boilerplate and functions, and then write logic myself.
- johnfn 3y agoI didn't want to say it in my initial post because I feared I would be downvoted for people thinking I was just spamming AI products. (I got downvoted anyways, so I guess it didn't matter.) I just use the Cursor IDE though, it's fabulous if you already use VSCode. And yeah I have no association with Cursor. Cursor is pretty smart about what to prompt GPT with. In the case that you described, you might prompt it with something like "check for prod" and I expect it would work. The reason it would have a better success rate is that it's usually quite good about what context to send over - it would likely pick up the type definitions (if you have those) or other files that do similar things and send those over as well. To give a full answer to your question, I actually wouldn't prompt it with "check for prod" at all because that would be underutilizing GPT. I would probably prompt at the complexity level of something like "edit this component so that if we're in prod, we show a warning message at the top". I would expect GPT to do that accurately for a component up to 100-200 lines virtually every time.